Single item inventory control under periodic review and a minimum order quantity

Transcription

1 Single item inventory control under periodic review and a minimum order quantity G. P. Kiesmüller, A.G. de Kok, S. Dabia Faculty of Technology Management, Technische Universiteit Eindhoven, P.O. Box 513, 5600 MB Eindhoven, The Netherlands This paper should not be quoted or referred to without the prior written permission of the authors

2 Single item inventory control under periodic review and a minimum order quantity Abstract In this paper we study a periodic review single item single stage inventory system with stochastic demand. In each time period the system must order none or at least as much as a minimum order quantity Q min. Since the optimal structure of an ordering policy with a minimum order quantity is complicated, we propose an easy-to-use policy, which we call (R, S, Q min ) policy. Assuming linear holding and backorder costs we determine the optimal numerical value of the level S using a Markov Chain approach. In addition, we derive simple news-vendor-type inequalities for near-optimal policy parameters, which can easily be implemented within spreadsheet applications. In a numerical study we compare our policy with others and test the performance of the approximation for three different demand distributions: Poisson, negative binomial, and a discretized version of the gamma distribution. Given the simplicity of the policy and its cost performance as well as the excellent performance of the approximation we advocate the application of the (R, S, Q min ) policy in practice. Keywords: Stochastic inventory model, Minimum order quantity, Markov Chain, Newsvendor inequality 1 Introduction Single echelon single location inventory models have been extensively studied in literature (see for an overview Silver et al. (3) or Zipkin (6)). Assuming linear holding and penalty costs, and fixed reordering costs, the optimality of (s, S) and (R, s, S)-policies in continuous review and periodic review, respectively, is proven. Because of their simple structure, these policies are widely applied in practice and have been implemented in many business information systems, such as ERP and APS systems. 1

3 However, inventory managers in practice are sometimes confronted with additional constraints and requirements. As an example we mention the situation at a globally operating packaged goods company, where process efficiency demands that production batches are at least of a minimum size. Other examples can be found in apparel industries, where a minimum order quantity is not uncommon, too (see also Fisher and Raman (1) and Robb and Silver (2)). The minimum order quantity restriction is not properly taken into account in the basic inventory models mentioned above. However, up to now little effort has been devoted to the modeling and analysis of inventory systems working with minimum order quantities. It has been proven that the optimal policy structure is complex (see Zhao and Katehakis (4)) and typically to complicated to implement in practice. Therefore, in literature the performance of different policy structures is investigated. For low periodic demand relative to the minimum order quantity a mathematical model is presented in Robb and Silver (2) to assist the decision maker when to order in case of a minimum order quantity. If the required amount is less than the minimum order quantity the actual ordersize can be increased or the order can be delayed. In a myopic approach both alternatives are compared in terms of costs in order to come up with formulae for the safety stock and the order threshold. In a large numerical study the authors show that their policy is outperforming a simple one, where the recommended order quantity is rounded up to the minimum amount. Fisher and Raman (1) have studied the stochastic inventory problem with a minimum order quantity for fashion goods. Since these products have very short life cycles with only few order opportunities, they investigate a two period model. They formulate a stochastic programming model to get insights in costs and the impact of the order constraint. A two parameter policy, called (R,s,t,Q min ) policy, is studied in Zhou et al. (5). It operates as follows. When the inventory position is lower than or equal to the reorder level s, an order is placed to raise the inventory position to s+q min. When the inventory position is above s but lower than threshold t, then exactly the required minimum amount is ordered. Otherwise no order is placed. In a numerical study the authors compare the proposed policy with the optimal one and conclude that the cost performance is 2

4 close to optimal. However, to compute the cost optimal (R,s,t,Q min ) policy the steadystate probability distribution of the inventory position is needed and the authors claim themselves that it is not clear how to calculate these steady state probabilities more efficiently than directly solving the linear system associated with the balance equations. Thus, searching for the optimal policy parameters is computational intensive. In this paper we propose a simple periodic review policy, called (R,S,Q min ) policy, where no order is placed as long as the inventory position, defined as the stock on-hand plus stock on-order minus backorders, is equal or larger than the level S. Otherwise an order is placed to raise the inventory to S. However, if this order is smaller than Q min we increase the order quantity to Q min. Note that this policy is a special case of the (R,s,t,Q min ) policy, viz. s = S Q min and t = S 1. Formulating the associated Markov chain model we can derive exact expressions for the holding and penalty costs for a given policy. This enables us to compute the optimal numerical value S opt for each given Q min. Since this procedure for finding S opt is computationally intensive, we develop simple news-vendortype inequalities from which a near optimal value S, can be routinely computed e.g. using an EXCEL spreadsheet. In a detailed numerical study we compare the performance of the proposed policy with an optimal (R,s,t,Q min ) and an optimal (R,s,S) policy with S s = Q min. Moreover, the performance of our approximation is tested, yielding to excellent results. We conclude that the simplicity of the policy and the expressions for the computation of the policy parameter as well as cost performance of the (R,S,Q min ) policy justify an implementation in practice. The remainder of the paper is organized as follows. In section 2, the model and the notation is introduced. In section 3 we first show how the optimal level S opt can be computed and afterwards we develop the news-vendor-type inequalities mentioned above. In section 4 an extensive numerical study is presented to test the performance of the policy and the approximation. Section 5 concludes the paper with a summary. 3

5 2 Model Description We consider a single item single echelon system with stochastic demand. In order to manage the inventory and place replenishment orders a periodic review system is used. We assume that the demand per period can be modeled with independent and identically distributed non-negative discrete random variables. Whenever demand cannot be satisfied directly from stock, demand is backordered. We further assume the length of the review period R to be given and without loss of generality we set it equal to one. Additionally, only order quantities of at least Q min units are permitted and we assume the value of Q min to be given. In order to determine replenishment times and quantities a so-called (R,S,Q min ) policy is applied. This policy operates as follows: at equidistant review timepoints the inventory position is monitored. If the inventory position is above the level S, then no order is triggered. In case the inventory position is below the level S, an amount is ordered which equals or exceeds Q min. An amount larger than Q min is only ordered, if the minimal ordersize Q min is not enough to raise the inventory position up to level S (see Figure 1 for an illustration of the policy). Inventory position S Q min Q min Q min S - X n R R R time Figure 1: The (R,S,Q min ) policy (leadtime equal to zero) The parameter S of the policy is therefore functioning as a reorder level as well as an order-up-to level. If the demand is always larger than the minimum order quantity, which may happen in case of small values of Q min, then the order constraint is not restrictive 4

6 anymore and the (R,S,Q min ) policy is similar to an order-up-to policy (R,S) with orderup-to level S. For large values of Q min the parameter S functions as a reorder level only, and the policy is equal to an (R,s,Q min ) policy with a reorder level s. In order to evaluate the inventory system the average costs per review period are considered, composed of two main components. On the one hand the company incurs inventory holding costs and on the other hand backorder costs arise from stockouts. An inventory holding cost h is charged for each unit in stock at the end of a period and a penalty cost b is charged for each unit short at the end of a period. Note that fixed ordering costs are not included in the cost model. The sequence of events is as follows. A possibly outstanding order arrives at the beginning of a period and the inventory position is reviewed and an order is placed if necessary. During the period, demand is realized and immediately satisfied if possible, otherwise demand is backlogged. Demand is satisfied according to a First-Come-First-Serve rule. At the end of the period holding and backorder costs are charged for each unit on stock or backordered. The aim of the paper is to analyze the (R,S,Q min ) policy and determine an optimal level S opt which minimizes the average holding and backorder costs per period in a stationary state, denoted as C(S). Let I + and I denote the stock on hand and backlog at the end of a period. Thus the objective function can be written as: C(S) = he[i + ] + be[i ] (1) In the remainder of this paper, the following notation will be used. 5

7 Q min S L D n D(i) q n X n Y n I h b E[X] σ(x) c v (X) X + X [a,b] : Minimum ordersize : Policy parameter : Leadtime : Demand during the nth period : Demand during i periods : The quantity ordered at the beginning of the nth period : The inventory position before ordering, at the beginning of the nth period : The inventory position after ordering, at the beginning of the nth period : Inventory level at the end of a period : Holding cost parameter per unit : Backorder cost parameter per unit : Expectation of a random variable X : Standard deviation of a random variable X : Coefficient of variation of a random variable X, (c v (X) := σ(x) : max(0,x) : max(0,-x) : The interval of integer numbers between a and b (a and b are also integers). E[X] ) 3 Computation of the optimal level S 3.1 Exact computation In order to determine the optimal level S opt, which minimizes the cost function given in (1), we use a Markov Chain {Y n }, where Y n is defined as the inventory position after ordering at the beginning of period n. This Markov Chain has the finite state space SS = [S,S + Q min 1]. The order quantity in period n + 1 follows from 0, Y n D n S q n+1 = Q min, S Q min < Y n D n < S (2) S Y n + D n, Y n D n S Q min 6

8 and the inventory balance equation is given as: Y n+1 = Y n + q n+1 D n (3) Hence, by replacing (2) in (3) we obtain: Y n D n, Y n D n S Y n+1 = Y n + Q min D n, S Q min < Y n D n < S (4) S, Y n D n S Q min Now we can easily compute the transition probabilities p ij = P(Y n+1 = j Y n = i) for all (i,j) SS 2 : { P(D n i S + Q min ) + P(D n = i S), j = S p ij = (5) P(D n = i j) + P(D n = i j + Q min ), j > S The steady state probabilities are defined as π i = lim n P(Y n = i) and they can be determined by computing the normalized eigenvector of the matrix P for the eigenvalue one: i SS π i = 1, 0 π i 1 π = P T π, π = (π i ) i SS (6) where (P = (p i,j ) (i,j) SS 2) denotes the matrix of the transition probabilities. Note that, for a given value of Q min and a given demand distribution, the steady state probabilities of the inventory position π i are independent from the numerical value of S. Conditioning on each possible value of the inventory position after ordering, we get for the objective function: S+Q min 1 C(S) = h i=s = (h + b) i (i j)p(d(l + 1) = j)π i + b j=0 S+Q min 1 i=s S+Q min 1 i=s (j i)p(d(l + 1) = j)π i ( i ) (i j)p(d(l + 1) = j) bi π i + be[d(l + 1)] (7) j=0 and we can easily evaluate the costs for given values of S. It is easy to show that the objective function is convex. Therefore, a bisection algorithm can be used for the optimization. 7 j=i

9 3.2 A near optimal level The time needed to find the optimal value S opt can be substantial, because we have to calculate the transition probabilities given in (5) and the system of linear equations (6) has to be solved. Moreover, a software package is needed enabling the determination of the steady-state probabilities numerically. However, due to the experience of the authors, many companies prefer to use spreadsheets instead of expert-only mathematical software packages. Therefore, we develop an easy-to-use alternative for the determination of the policy parameter S. Our starting point is a periodic order-up-to policy without order size restrictions. In such a situation it is well-known that the optimal order-up-to level is given as the smallest value of S satisfying the following news-vendor-type inequality P(D(L + 1) S) c u c u + c o (8) where c u is denoting the underage costs and c o the overage costs per unit. In the sequel we will show how expression (8) can be adapted such that a near optimal value of S can be determined for the (R,S,Q min ) policy. Our new expression is based on two observations: Formula (8) has to be adapted to cover the fact that replenishment cycles may start with different inventory positions For two special cases, namely Q min = 1 and Q min very large, it is known that the parameter S can be computed by news-vendor inequalities According to the last observation and the completely different behavior of the policy for small and large values of Q min we could not find one formula for all possible instances. We propose to choose the maximum of two values S 1 and S 2, where the first one is appropriate for small values of Q min and the latter one for large values. We define the near optimal value S as follows: S = max{s 1,S 2 } (9) In order to come up with a news-vendor-type inequality for S 2 we first consider the limiting case where Q min much larger than period demand E[D]. Then the policy is similar to an (R,s,Q min ) policy and it is known that the inventory position after ordering is uniformly 8

10 distributed on [S,S + Q min 1]. In order to compute S 2 we approximate the steady state distribution of the inventory position with a uniform distribution and we adapt (8) conditioning on the starting inventory position. Then the numerical value of S 2 is determined as the smallest value of S satisfying: 1 Q min Q min 1 k=0 P(D(L + 1) S + k) b b + h (10) In order to derive a formula for S 1, which performs better in case of small values of Q min, we use a different approach. Since the distribution of the inventory position is not known in this case, we discuss how to approximate the overage and underage costs. Any underage of a single unit will result in a backorder penalty b assuming that a backorder will only last one period. However, the chance that an overage will last more than one period is not negligible. An overage can only be compensated if an order is placed in the following period, which means, the inventory position must be below the level S. But due to the minimum order quantity the inventory position can also be above S and it is impossible to adjust the decision. Thus, in order to approximate the overage costs we have to take into account that an overage can last for several periods. We assume that the number of periods we have to wait until the overage can be compensated, has a geometric distribution with parameter p. Since the probability p of this distribution is not only dependent on the demand but also on the starting inventory level we approximate p, ignoring the latter effect, and we use p = P(D Q min ) (11) As a result, the overage costs can be estimated as c o = h ip(d Q min ) i 1 (1 P(D Q min )) = i=1 h 1 P(D Q min ) (12) A numerical value for S 1 is determined as the smallest value of S satisfying the following inequality. P(D(L + 1) S) b + b h 1 P(D Q min ) (13) 9

11 4 Numerical study The objective of this section is two-fold. Firstly, we are interested in the performance of the (R,S,Q min ) policy compared to other policies. Secondly, we test the performance of the (R,S,Q min ) policy compared to the (R,S opt,q min ) policy. We consider three different discrete demand distributions: Poisson distribution Negative binomial distribution A discretized version of the gamma distribution as defined below P(D = 0) := F(0.5) P(D = i) := F(i + 0.5) F(i 0.5), i = 1, 2,...,D max P(D = D max ) := 1 F(D max 0.5) where F(x) denotes the cumulative distribution function of a gamma distributed random variable. For each demand scenario we set up a factorial design to cover a wide range of conditions. We introduce a non-dimensional parameter m = Q min /E[D] as one factor and consider five levels of this factor. We expect for numerical values of Q min close to the mean period demand the (R,S,Q min ) policy differs most from the order-up-to policy (R,S) or the fixed order size policy (R,s,Q). Therefore, we choose m {0.5, 0.9, 1.0, 1.1, 1.5}. Additionally, we consider the factors leadtime L, inventory holding cost h, mean demand E[D], and coefficient of variation of demand c v (D) (in case of a Poisson distribution we can omit this factor). For each of the latter factors we employ three levels: low, middle, and high. Further, for all examples the backordering cost parameter p = 100 is fixed. The numerical values for the input parameters are given in Table 1: 10

12 Poisson Negative binomial, Discretized Gamma Factor level level Factor low middle high low middle high L L h h c v (D) c v (D) E[D] E[D] m m Table 1: Numerical values of the input parameters 4.1 Performance of the policies In order to illustrate the economic impact of the (R,S,Q min ) policy we compare it with two other policies. The first alternative is a policy often used in practice and called min-max policy. Formally the policy is an (R,s,S)-policy, i.e. a periodic (s,s) policy (see Silver et al. (3)). Due to the minimum order quantity the policy parameters are restricted to S s = Q min. The other policy is the two-parameter policy (R,s,t,Q min ) (s t < s + Q min ) proposed by Zhou et al. (5). If X n denotes the inventory position in period n before ordering, then the replenishment decision can be described as follows: 0, X n > t q n = Q min, s < X n t (14) s + Q min X n, X n s The optimal policy parameters can also be determined by modeling the system as a Markov Chain (for more information see Zhou et al. (5)). Note that the (R,S,Q min ) policy is a special case of the (R,s,t,Q min ) policy where s = S Q min and t = S 1. Therefore, the average minimal costs of the (R,S,Q min ) policy cannot be lower than the minimal costs of the (R,s,t,Q min ) policy. Still, we are interested in the cost difference, because in case of small deviations the simplicity of the (R,S,Q min ) policy will justify its application, given the computational complexity of finding the optimal parameters of the (R,s,t,Q min ) policy. For all three policies we compute the optimal policy parameters and the minimal costs. We denote the minimal costs for the (R,S,Q min ) policy by C S, for the (R,s,t,Q min ) 11

13 policy by C s,t, and for the (s,s+q min ) policy by C s,qmin. For each example i we compute the relative deviation in percentage: i (s,t) = C S C s,t C s,t 100% and i (s,q min ) = C S C s,qmin C s,qmin 100% (15) We also compute the average cost deviation over N examples as: av = 1 N N i (16) i=1 as well as the minimal and the maximal cost deviations min = min i i, max = max i i (17) The numerical results of the 135 examples for Poisson distributed demand are given in Table 2. Factor Level (R,s,t,Q min ) (R,s,s + Q min ) av (s,t) max (s,t) min (s,q min ) av (s,q min ) max (s,q min ) L h E[D] m Table 2: Comparison of policies (Poisson distribution) It can be seen that on average the two parameter policy is not much more cost effective than our proposed one parameter policy. The largest deviations occur when the minimum order quantity is equal to the average demand. 12

14 Although the performance of the (R,s,s+Q min ) policy is for some examples close to our policy, the cost differences are small and on average the (R,S,Q min ) policy leads to lower cost. Moreover, there are also some examples where really large costs differences can be observed. In Table 3 the results of the 405 examples with a negative binomial distribution are depicted. Factor Level (R,s,t,Q min ) (R,s,s + Q min ) av (s,t) max (s,t) min (s,q min ) av (s,q min ) max (s,q min ) L h c v (D) E[D] m Table 3: Comparison of policies (Negative Binomial) On average the cost improvement of an (R,s,t,Q min ) policy compared to our (R,S,Q min ) policy is small. And also in the worst case the cost deviation is smaller than 4%. It can further be seen that the differences are getting smaller with increasing demand variability. This is also the reason why the performance in case of negative binomial demand is better than in case of Poisson demand. Although the (R,s,s + Q min ) policy performs in some situations slightly better, this policy cannot be advocated since on average the costs are much larger and in the worst case the differences can be really large. 13

15 In Table 4 the results of the discretized gamma distribution are shown. Factor Level (R,s,t,Q min ) (R,s,s + Q min ) av (s,t) max (s,t) min (s,q min ) av (s,q min ) max (s,q min ) L h c v (D) E[D] m Table 4: Comparison of policies (Discretized gamma distribution) Similar to the case of negative binomial distributed demand, the cost difference between the two parameter policy (R,s,t,Q min ) and the one parameter policy (R,S,Q min ) is so small that the simplicity of the policy as well as the computational effort for computing the policy parameters, justify the application of the simpler policy. Again, the min-max policy is not a good alternative. 4.2 Performance of the approximation In order to measure the accuracy of the approximation we have computed the optimal level S opt using the Mark Chain model described in section three and we have compute S as given in (9). For both policy parameters we compute the average costs and the 14

16 relative difference of them as follows: δ := C(S ) C(S opt ) C(S opt ) 100% (18) Average, minimum and maximum deviations are defined similar as above. The results of the 135 examples under Poisson distributed demand are summarized in Table 5. Additionally, it is interesting to mention that in 62% of the examples the heuristic has computed the optimal solution and in 87% of the examples the relative deviation of the costs was smaller than 1%, which is really excellent. Factor relative cost deviation Level δ min δ av δ max L h E[D] m Table 5: Performance of the approximation (Poisson distribution) The largest differences are observed for minimal order quantities close to the average period demand. Moreover, with increasing demand uncertainty the performance of the approximation is improving. This effect can also be seen for the other demand distributions (see Table 6). For the discretized gamma distribution the performance of the approximation is increasing with increasing minimal order quantity. With an overall performance in case of a negative binomial (discretized gamma) distribution with an average relative deviation of 0.28% (0.24%), a maximum relative deviation of 2.46% (4.83%), and 15

17 with 93% (89%) of the examples with a cost deviation smaller than 1% we can conclude that the simple expressions for the computation of S lead to an excellent cost performance within the class of the (R,S,Q min ) policies. relative cost deviation Factor Level negative binomial adapted gamma δ min δ av δ max δ min δ av δ max L h c v (D) E(D) m Table 6: Performance of the approximation 5 Summary and Conclusions In this paper we have studied an easy-to-implement replenishment policy for a stochastic inventory system where order quantities are required to have a minimum size. In an extensive numerical study we have illustrated that the cost performance of this one parameter policy is close to the cost performance of a more sophisticated two-parameter policy. Moreover, we provide simple and easy-to-use formulae to compute near optimal policy parameters. Since the financial implications of the more complex policy are not significant and also optimal parameter calculations are much more cumbersome for this policy, we advocate the application of our policy in practice. 16

18 References [1] Fisher M, Raman EA, (1994) Reducing the cost of demand uncertainty through accurate response to early sales. Operationa Research 44, [2] Robb DJ, Silver EA, (1998) Inventory management with periodic ordering and minimum order quantities. Journal of the Operational Research Society 49, [3] Silver EA, Pyke DF, Peterson R, (1998) Inventory Management and Production Planning and Scheduling. Third Edn., John Wiley & Sons: New York. [4] Zhao Y, Katehakis, MN (2006) On the structure of optimal ordering policies for stochastic inventory systems with a minimum order quantity. Probability in the Engineering and Informational Sciences 20, [5] Zhou B, Zhao Y, Katehakis, MN (2007) Effective control policies for stochastic inventory systems with a minimum order quantity and linear costs. International Journal of Production Economics 106, [6] Zipkin, P.H. (2000) Foundations of inventory management. First Edn., McGraw-Hill Companies. 17